Here is every byte that leaves your machine, for each tool

Does Wispr Flow take screenshots?

Not any more, by its own help page. Wispr Flow's Context Awareness page, read on 7 October 2026, says the feature "works from text and does not take screenshots"; in September the same page listed a screenshot among the context that could go with a dictation. It still reads text and sends it: with Context Awareness on, which is the default, the text in the box before, selected and after your cursor, on-screen text, file and variable names in coding apps, your user identifier in the app, the apps in your session and, on a Mac, conversation history. Switched off, Flow still sends the app's name and the website's domain.

What are the privacy concerns with Wispr Flow?

Its own pages name them. Transcription always happens in the cloud, so the audio of every dictation leaves your machine, and there is no local mode: offline, Flow saves the audio to retry later. The service is not end-to-end encrypted, because it must process the audio, and it stores customer data in the United States. Context Awareness sends on-screen text by default. "Improve the model for everyone", which lets your audio, transcripts and edits train its models, is on by default for Free and Pro, and off and locked on Enterprise; on Growth its pricing page gives an admin control over it. And with training off, its data-sharing page says Flow still uploads dictation metadata such as word count, app, URL and timestamp. Deixis sends none of that: no audio, no screen text, no telemetry.

Does Deixis upload my audio?

Not from a dictation, under any setting or on any plan. Speech recognition runs inside the app on your own device, whichever of Windows, Mac, Android or the iPhone keyboard in beta it is, so the recording, the model and the transcript stay there. A meeting is recognised on the device too, and its audio stays there unless you choose otherwise, below. The voiceprints meeting mode keeps to tell speakers apart stay on the machine whatever you choose: they are never uploaded, and can be deleted per person.

The one setting that sends a meeting's audio is Deixis cloud processing for meetings on Android, in Settings › Meetings. It is off until you turn it on, and it comes with Pro and the trial, up to 20 hours of meetings a month. With it on, once a meeting ends the phone sends its audio through Deixis to Whisper on Cloudflare Workers AI in pieces, each its own request, with your vocabulary and any names or words you gave Deixis for that meeting, and the text comes back to the phone. It sends on Wi-Fi: on mobile data a meeting waits for Wi-Fi unless you tap Send now on mobile data on it, or turn on Use mobile data for Deixis cloud, which is off until you do. Deixis stores and logs none of the audio, the text or those words — they travel in the body of each request, which our logs never hold — and Cloudflare does not train on them. Deixis keeps one number, how many seconds of audio your account has had processed that month, and its service log keeps one line per piece: your account, whether it worked, how many seconds it was and how long each step took, never what was said. Past the allowance the phone processes the meeting itself; a meeting that reaches the allowance part-way has sent some of its pieces before the phone takes over. Apart from that, the one file the app can carry audio off the machine in is a screen recording you narrate and then choose to share with Share a link, which uploads that one file and needs a Deixis account.

One recording can leave without the app, and the table names it: a website's Deixis mic, pressed in a phone's browser, or on a computer where Deixis is not installed, not running or not yet connected to that site. Then your browser records that one take and sends it through Deixis to Whisper on Cloudflare Workers AI, with the website's origin and your IP address, as any web request carries them. The text comes back, Deixis stores and logs neither the recording nor the text, Cloudflare does not train on them, and the line beside the mic says so on every take.

What does Deixis send when cleanup is on, or when you start the trial?

When you configure a cleanup endpoint and an utterance escalates to it, Deixis sends the transcript, your vocabulary terms, and where the text is going: the focused app's name, the control and field name, and your previous utterance in that field. It never sends the window title, which is the most identifying thing in that structure. With no endpoint set, nothing is sent and the local rules engine does the cleanup. The Deixis endpoint asks its model router to use only zero-data-retention endpoints whose policy does not collect inputs, so prompts are not stored or trained on; the privacy policy names every party in that chain.

Starting the trial sends your email address, so the sign-in code has somewhere to go, and, once, the partner whose download link you used and the first partner website that connected to Deixis on this computer, if there are any, so that partner can be credited. Without a trial none of this is sent.

Can I see what a tool sends without trusting its privacy policy?

Yes, and it is worth doing. Put a proxy such as mitmproxy between the app and the network, or watch outbound connections with a firewall that logs them, then dictate a sentence. A tool that sends audio produces a large upload on every utterance; a tool that sends screenshots produces periodic image-sized posts. The tables here are what the documentation says; a proxy shows you what the software does.

Why do some rows say "not documented", and how often are they re-checked?

Because the vendor's privacy policy or documentation did not address that row when we read it, on the date shown. We do not guess in either direction; if you know of a page that answers it, send it to support@deixis-voice.com and the row gets updated with the date. Every entry carries the date it was checked, and a fact older than about three months raises a warning in our build so it gets re-read, because vendors change their pricing and their privacy terms without announcing it.